OpenAI Realtime API alternatives: decide which layer you are replacing

Realtime is one model doing speech, reasoning, tools and speech output in one session. Most alternatives replace only part of that, so start by naming the part.

Duplex Compute · October 13, 2026 · Rates and versions checked October 13, 2026

The Realtime API uses one model to interpret audio, decide what to do and respond in speech. Its current model, gpt-realtime-2.1, supports function calling, configurable reasoning effort and image input. It also handles barge-in: with voice activity detection on, it cancels the current response when the user starts talking. Neither feature is missing, so neither is a reason to switch on its own.

The usual reasons are narrower. Realtime bills by tokens, at $32 per million audio input tokens and $64 per million audio output tokens, and resends the conversation to the model on every response, so long calls get expensive in ways that are hard to predict. A session lasts at most 60 minutes. The reasoning model is fixed to OpenAI's speech model. Some teams want open weights or their own hardware.

Six kinds of replacement

Gemini Live

Google's Live API is the closest like-for-like swap. Its pricing section for Gemini 3.8 Live, 3.8 Live Extended Thinking and 3.1 Flash Live Preview lists audio input at $3 per million tokens (about $0.005 per minute) and audio output at $12 per million (about $0.018 per minute). Tools use function declarations. Pin the model before you design around async tools: the capability table on the tools page marks Gemini 3.1 Flash Live Preview as synchronous only and Gemini 2.5 Flash Live Preview as synchronous and asynchronous. It does not cover 3.8.

Session lifetime works differently. Without context compression, audio-only sessions stop at 15 minutes, and a single connection lasts around 10 minutes. Session resumption and compression extend a conversation across connections. Your client has to handle the reconnect.

GPT-Live and other voice frontends

GPT-Live is OpenAI's other voice API, and a different architecture. It listens while speaking and delegates reasoning and tools to a backend you choose, including an agent you already run on another provider. It bills $0.05 per session minute, with backend usage on top. Our GPT-Live vs. Realtime note compares the two in detail.

Deepgram Voice Agent

Deepgram's Voice Agent API runs listening, thinking and speaking over one connection, and its function calls can run in your client or on Deepgram's side. The Standard pay-as-you-go tier is $0.075 per minute of WebSocket connection time. Cheaper tiers exist where you bring your own LLM or TTS, and those providers then bill you separately.

Pipecat and LiveKit Agents

Frameworks orchestrate models that other providers supply. LiveKit Agents runs Python or Node.js agents as participants in LiveKit rooms, under Apache 2.0, with plugins for most major AI providers. Pipecat is an open-source Python framework, and it can run over LiveKit's WebRTC transport, so the two are not exclusive. Either can wrap a speech-to-speech model such as Realtime itself, or chain separate speech-to-text, LLM and text-to-speech providers.

Hosting is a separate meter from the models. Pipecat Cloud's agent-1x profile costs $0.01 per active minute. LiveKit Cloud includes agent-session minutes by plan, then charges $0.01 per minute. Neither figure includes the model calls.

Open full-duplex models

Moshi models two audio streams, the user's and its own, and ships PyTorch, MLX and Rust implementations. Its weights are CC-BY 4.0. Its README documents no tool calling. NemotronLabs VoiceChat is an 11B full-duplex model with tool calling, and its README is frank about limits. The documented deployment needs an NVIDIA GPU with at least 80 GB of memory. NVIDIA recommends at most five tools per session and says the user cannot interrupt the agent while a tool runs.

A hosted wrapper can expose less than the checkpoint. .wave's Nemotron VoiceChat beta costs an introductory $0.005 per minute, caps sessions at 120 seconds and supports no tools, custom instructions or text input.

Vapi and Retell

These are platforms for running phone agents, with numbers, transfers and dashboards. They sit on top of model providers rather than replacing them. Vapi charges $0.05 per minute for hosting and passes model costs through at cost. Retell lists $0.055 per minute for voice infrastructure plus your chosen TTS and LLM, and quotes $0.07 to $0.31 per minute depending on configuration. Choose one if you need the phone operations. A web or app developer may not.

Requirements to check before switching

No row in this table is a complete bill. Per-minute rates meter different things: open connections, generated audio, hosting containers. The voice agent pricing note has the formula for each billing model, so you can put two options on the same workload.

A migration checklist

1. Record a set of real calls on Realtime today, with tool calls, corrections, interruptions and the final application state.
2. List the Realtime features your code depends on: semantic_vad, image input, truncation on interruption, the function_call_output loop.
3. Map each to the candidate's equivalent, or note that your code must now do it.
4. Decide what happens to a tool call already running when the user interrupts or corrects it.
5. Rebuild the bill from logged usage: connected time, speaking time, tokens, cache hits and hosting.
6. Replay the same scenarios on the candidate and compare task success as well as speed and cost.

Step six matters most. A provider that answers faster and books the wrong date costs more per finished task. OpenAI's voice agent evaluation cookbook shows how to check tool calls and resulting state, not only the audio.

Choosing by application

Duplex Compute belongs in the first row. It is a full-duplex voice API with tool calling, configurable behavior and long conversations, priced at $0.01 per connected minute. That is the voice-session rate; your backend model and tools cost what they cost.

Related notes

• GPT-Live vs. Realtime API
• Voice agent pricing: calculator and billing formulas
• Pipecat vs. LiveKit
• Gemini Live API pricing
• Self-hosted voice AI vs. managed APIs